Predicting DOGE AI Deregulatory Tool Recommendations
Last month, I learned about a planned DOGE AI deregulation tool (called SweetRex) that would be applied to review agency regulations and guidance to find what could be cut if it wasn't directly tied to what Congress said they wanted in the original language of the law.
This is all coming against the backdrop of 1) the Administration aiming to repeal regulations that aren't explicitly grounded in statutory text (per Executive Order 14219, issued April 2025), and 2) the Supreme Court's Loper Bright decision that removed “Chevron deference”—meaning if Congress didn't explicitly say it in the statute, agencies can't assume they have the authority to regulate it.
So, what I decided to do was take what I thought would be the AI prompt and ask multiple generative AI tools to review an agency regulation that I know well - the 2008 Clean Water Act Section 404 compensatory mitigation rules. These are the regulations that govern how wetland and stream impacts get offset. .
The TLDR: AI analysis flagged roughly half the mitigation rule as potentially vulnerable, but these were mostly nuanced implementation details like whether agencies should consider watershed approaches when siting mitigation projects, and how long monitoring should last. What’s more fundamental and at risk is the concept of the mitigation hierarchy (first avoid impacts to wetlands, then minimize, then compensate) and mitigation preference hierarchy (use mitigation credits that represent restoration that has already achieved ecological performance standards over restoration that has not yet begun). This isn’t bad regulation, it’s just not explicitly tied to what Congress specifically asked for in legislation.
Where the models agree are where provisions have explicit statutory hooks (like performance standards and mitigation banking frameworks mandated by the 2003 National Defense Authorization Act, or NDAA §314(b)) - these are on solid ground.
What I Did
I developed a standard prompt that was our best approximation of the deregulatory process/prompt described in an August 14th Wired article and associated leaked powerpoint presentation. The regulation/statutes we gave in the prompt include the following:
40 CFR part 230, Subpart J (aka 2008 Rule, p.95 - 113)
Clean Water Act section 404 permitting regulations / US Code §1344, Permits for dredged or fill material
Public Law 108–136—NOV. 24, 2003, National Defense Authorization Act (NDAA) for Fiscal Year 2004 (herein NDAA 2003), §314(b).
2016 Congressional Research Service (CSR) report “Clean Water Act: A Summary of the Law” which describes the original enactment and subsequent major amendments (e.g., indicating Congressional intent)
LoperBright commentary blog (Sixth Circuit Appellate Blog, July 2024)
EO 14219, April 2025, Directing the Repeal of Unlawful Regulations
I used the prompt with four GenAI models: Gemini (the article describes the AI model being based primarily on Gemini), Claude, ChatGPT, and Perplexity. I then loaded the output documents - which did not necessarily return results in the specific format suggested - into Claude and asked it to organize it into a close approximation of what the powerpoint presentation showed, pointing out where the models’ arguments agreed or diverged.
But I didn't just ask the models to identify deregulation targets. I also asked them to provide counter-arguments from a neutral-to-slightly-pro-regulation perspective—essentially, what would the defense look like if it were grounded in congressional intent? This dual approach let me see both where the vulnerabilities lie and where agencies might have defensible statutory interpretations.
I focused specifically on Clean Water Act Section 404 mitigation regulations (40 CFR Part 230, Subpart J) because it's an area where I have deep subject matter expertise and can actually fact-check the AI's statutory interpretations (by the way, the models didn’t return something totally off base, usually just nuanced misunderstanding). The analysis covered all eight regulatory sections, from purpose and definitions through mitigation banking requirements.
What I Found
The AI models behaved differently but reached similar conclusions about vulnerability. Gemini provided the most aggressive deregulatory analysis, relying heavily on Loper Bright to argue that agencies can no longer fill statutory gaps with policy judgments. Claude and ChatGPT showed similar, slightly more moderate approaches. Perplexity only analyzed the first two sections before stopping (I could perhaps have prompted it more but using Perplexity was a pilot test for me and… I’m not married to it let’s just say).
The safest provisions were those with explicit congressional mandates:
§230.95 (Performance Standards) – All models agreed this is directly required by the 2003 NDAA
§230.98 (Mitigation Banks) – Strong consensus due to explicit statutory language about mitigation banking regulations
The most vulnerable provisions lacked clear statutory foundation:
§230.93 (General Requirements) – The mitigation hierarchy (preference for mitigation banks over permittee-responsible mitigation) and watershed approach were flagged as agency interpretation rather than congressional mandate
§230.97 (Management) – "In perpetuity" protection requirements appear nowhere in the statute, though the counter-argument notes that permanent impacts logically require permanent compensation
§230.96 (Monitoring) – The five-year minimum monitoring period was questioned since Congress never specified timeframes, though the defense pointed to ecological science on wetland establishment timelines
What came out of this exercise was that even where there are longstanding, science-based requirements that are more operational than foundational concepts - these are vulnerable if Congress didn’t explicitly authorize them. Things like "adaptive management," specific mitigation ratios, and financial assurance requirements were flagged as potentially exceeding statutory scope. But the counter-arguments consistently pointed to congressional intent signals: oversight hearings that identified "paper mitigation" problems, funding decisions that endorsed these practices, and the 2003 NDAA's requirement for "equivalent standards" that agencies argued necessitated detailed implementation frameworks. Going forward, Congress may need to write out all of these details rather than relying on agencies to make reasonable interpretive choices. Yes, we’ll need a Congress full of subject matter experts in a variety of fields (or we’ll need lobbyists?).
We’re in a new environment post-Loper Bright and we don’t have enough experience to know how picky the Administration, Congress, or the Supreme Court will be going forward. It’s heartening that agency staff - presumably those with great subject matter expertise - will have the chance to review these suggested deletions and provide feedback.
Why This Matters
While DOGE may be the first to use AI to streamline regulations (at least that we’ve seen), it likely won’t be the last. In fact, Virginia plans to test both an AI-powered regulatory review (coming on the heels of a manual review already conducted by agency staff - so the contrast should be interesting), as well as an AI-enhanced "proof of concept" permit review that would automatically assess application completeness, check regulatory compliance, and draft initial permits for agency review. The pilot will first be applied to the Virginia Water Protection Program permit review process and is anticipated to be deployed around September 2026.
What an AI Analysis Can—and Can't—Tell Us
Will DOGE's actual analysis look like mine? I don’t know. But we'll find out. The good news is that AI tools are pretty consistent in identifying the same statutory gaps, but they can't replace legal expertise or policy judgment. My analysis benefited from subject matter knowledge that let me fact-check the AI's statutory interpretations and spot places where models missed nuance or context.
The real test comes when DOGE releases their agency deletion plans. According to their timeline, agencies were supposed to finalize plans in August 2025, with Notice of Proposed Rulemakings going through OIRA in September and October, followed by expedited comment periods in November and December. Note that we have not heard of anything coming out officially.
DIY AI Deregulation Predictions?
I ran this analysis on regulations I know intimately. The methodology could be replicated for any federal regulation. Contact me if you'd like to view my internal report and full prompts.
